GPU neoclouds 2026: Pricing, contracted power and which provider fits your workload

Best GPU neoclouds, 2026: pricing, contracted power, and who fits which workload

CoreWeave’s Q2 2026 revenue of $2.575 billion (+112% year‑over‑year) is a blunt signal. The GPU cloud market has moved from niche to an industrial sector with specialized vendors, heavy capital plans, and bespoke contract mechanics (figures checked Aug 21, 2026).

Public rate cards tell part of the story. The rest lives in anchor contracts, prepayments, auctions, and sales‑led SKU access. Below I compare five GPU‑focused “neoclouds” (CoreWeave, Nebius, Lambda, Crusoe, and Groq) on buyer‑relevant axes: published list prices, spot/preemptible tiers, contracted or deployed power, hardware roadmap access, and independent cluster ratings. Where the ledger gives specific numbers I flag the source and date so you can test them in an RFP.

Methodology and definitions (short and usable)

Signals used: published on‑demand pricing (per‑GPU/hour where available), published spot/preemptible pricing, deployed/contracted data‑center power (MW/GW), hardware roadmap access (next‑gen silicon, NVL72/GB300), and SemiAnalysis ClusterMAX 2.0 independent ratings.
Data cut‑off: figures checked Aug 21, 2026. Primary sources are vendor earnings releases, pricing pages, SEC 6‑K filings, and the SemiAnalysis ClusterMAX 2.0 report (Nov 2025).
Per‑GPU derivations: when vendors publish node prices only, the per‑GPU figure below is the node‑price divided by the published GPU count for that node (per‑node price ÷ GPUs per node). That is the standard conversion used in the market, and rounding and any software/instance overheads are noted where the vendor provides them.
Terms explained: NVL72 = NVIDIA Vera Rubin NVL72 rack system. H100/H200 = NVIDIA Hopper/H200 generations. B200/B300 = NVIDIA Blackwell classes. MI300X/MI355X = AMD Instinct classes. LPU = Groq lightweight processing unit (inference chip).

At a glance, the five vendors (key, scannable facts)

CoreWeave, Platinum ClusterMAX (SemiAnalysis), Q2 revenue $2.575B, active power ~1.5 GW, contracted ~3.7-4.2 GW, validated NVIDIA Vera Rubin NVL72 in Q2 2026. Published H100 per‑GPU (derived): $6.16/GPU‑hr; spot H100 NA: $2.46/GPU‑hr. (CoreWeave Q2’26 release, CoreWeave pricing, SemiAnalysis ClusterMAX 2.0.)
Nebius, Gold ClusterMAX (SemiAnalysis), Q2 group revenue $582.3M; AI cloud revenue $574.9M; AI cloud ARR $3.0B; target contracted power 5 GW by year‑end 2026. Published H100 (HGX): $3.85/GPU‑hr; preemptible H100: $2.15/GPU‑hr; published B300: $7.85/GPU‑hr. (Nebius Q2’26 letter, pricing page, Form 6‑K.)
Crusoe, Gold ClusterMAX (SemiAnalysis), contracted AI infra 4.9 GW across five U.S. campuses; development pipeline >40 GW. Lists AMD MI300X on rate card. Published H100: $3.90/GPU‑hr; cheapest H200: $4.29/GPU‑hr. (Crusoe announcements, pricing page, SemiAnalysis.)
Lambda, Silver ClusterMAX (SemiAnalysis), Series E >$1.5B (Nov 2025), $1.0B credit facility (May 2026), Microsoft multibillion agreement (Nov 3, 2025). Published B200 lowest in set: $6.69/GPU‑hr; H100 SXM 8x: $3.99/GPU‑hr; no published spot tier. (Lambda pricing and press releases.)
Groq, inference specialist (GroqCloud), licensed inference IP to NVIDIA (reported Dec 2025), joined NVIDIA Cloud Partner Aug 12, 2026. Groq runs per‑token LPU endpoints (no GPU‑hr pricing). Capacity: 13 DCs, 54 MW now; plans 200+ MW in 2027. (Groq newsroom; media reporting.)

Ranking and what matters (short profiles with sourcing)

CoreWeave, premium, validated systems and the highest independent rating

Why it ranks here: CoreWeave is the only Platinum provider in SemiAnalysis’s ClusterMAX 2.0 (reported twice in ClusterMAX), and it completed a public bring‑up and validation of NVIDIA’s Vera Rubin NVL72 in Q2 2026 (CoreWeave Q2’26 release, SemiAnalysis ClusterMAX 2.0).

Key numbers (checked Aug 21, 2026). Q2 revenue $2.575B (+112% YoY). Revenue backlog roughly $104B with more than $25B of new commitments added early Q3. Active power about 1.5 GW. Contracted power reported around 3.7-4.2 GW (CoreWeave DC page also lists 4.2 GW+ across 51 DCs). Management guidance: full‑year 2026 revenue $12.4-13.2B and capex guide $35-39B (CoreWeave Q2’26 release and data pages).

Published pricing: H100 per‑GPU (derived from node price): $6.16/GPU‑hr. Spot H100 (North America): $2.46/GPU‑hr. Reserved discounts up to 60% (CoreWeave pricing page).

When to pick CoreWeave: you need validated NVL72/GB300 class systems, enterprise SLAs, and operational maturity. Expect to pay a premium but get public engineering evidence and a full rate card.
Nebius, headline low H100 list price and contract scale

Why it ranks here: Nebius publishes the cheapest H100 (HGX) on‑demand list price in this set and is the only provider here to publish a B300 on‑demand price. Nebius is aggressively contracted with large anchor deals and a cash‑heavy balance sheet for buildouts (Nebius Q2’26 results, Form 6‑K filings).

Key numbers. Q2 group revenue $582.3M (+454% YoY). AI cloud revenue $574.9M (+514% YoY). AI cloud ARR $3.0B. AI cloud adjusted EBITDA margin 49.7%. Ended Q2 with $8.0B cash. Contracted power target raised to 5 GW by year‑end 2026. Reported running capacity auctions and structuring large prepaid deals (Nebius Q2 letter, Form 6‑K, investor materials).

Published pricing: H100 (HGX) on‑demand $3.85/GPU‑hr. B200 $7.15/GPU‑hr. B300 on‑demand $7.85/GPU‑hr (only public B300 here). Preemptible H100 $2.15/GPU‑hr (preemptible about 45% below on‑demand per Nebius pricing page).

When to pick Nebius: you prioritize headline list pricing for H100 or B300 and are prepared to negotiate contract terms such as prepayments and utilization floors that change realized economics. Nebius’ public examples show large deals averaging more than $1B with roughly 70% prepaid and structural terms that create utilization floors (Nebius Form 6‑K and investor letter).
Crusoe, campus builds and AMD capacity on a public rate card

Why it ranks here: Crusoe’s commercial model pairs data‑center campus projects (Abilene) with published AMD Instinct pricing. It reports contracted AI infrastructure of 4.9 GW across five U.S. campuses and a development pipeline greater than 40 GW (Crusoe announcements).

Published pricing: H100 per‑GPU about $3.90/GPU‑hr. Cheapest H200 $4.29/GPU‑hr. Crusoe is the only provider in this set that lists AMD MI300X on its rate card (MI355X available via sales) (Crusoe pricing page).

When to pick Crusoe: you want AMD evaluation in production or campus‑scale capacity via long‑term build partners (Oracle/OpenAI Abilene site; Microsoft campus announcements are sourced in Crusoe press coverage).
Lambda, mid‑market clusters, Microsoft anchor, predictable on‑demand pricing

Why it ranks here: Lambda targets mid‑market clusters and self‑serve customers. It has a multibillion, multi‑year agreement with Microsoft (Nov 3, 2025 announcement), a Series E greater than $1.5B (Nov 2025), and a $1.0B senior secured credit facility (May 2026) (Lambda press releases).

Published pricing: lowest B200 on‑demand instance in this set at $6.69/GPU‑hr. H100 SXM 8x instance $3.99/GPU‑hr. Lambda publishes no spot/preemptible tier (Lambda pricing page). SemiAnalysis ClusterMAX rates Lambda Silver.

When to pick Lambda: you want straightforward, self‑serve on‑demand clusters for mid‑size training and prefer a vendor with a Microsoft anchor rather than heavy sales‑led enterprise contract mechanics.
Groq, LPU inference specialist now a hybrid cloud partner

Why it ranks here: Groq built GroqCloud for token‑priced inference (LPU endpoints) and materially changed its market position after NVIDIA licensed Groq’s inference technology in Dec 2025 (reported by media). Groq joined NVIDIA’s Cloud Partner program Aug 12, 2026 and says it plans to bring NVIDIA accelerated computing online “in the future” (Groq newsroom, announcement). Groq’s pricing model is per‑token rather than per‑GPU‑hour.

Key numbers: capacity 13 DCs / 54 MW now. Plans 200+ MW in 2027. Raised $650M in June 2026 and $350M Series A on Aug 17, 2026 at a $3.5B valuation (Groq press announcements, fundraising reports).

When to pick Groq: production, high‑volume, low‑latency inference where per‑token economics and an LPU stack matter. If you plan to mix Groq inference with NVIDIA GPU workloads, validate the timing and fidelity of Groq’s promised NVIDIA GPU availability.

What the published prices actually mean, and what they don’t

Public list prices are a baseline, not the clearing market price. Large enterprise deals commonly use prepayments, capacity reservations, or structured purchase obligations that materially change realized unit economics (Nebius examples show more than $1B deals with about 70% prepayment and clause structures that function as utilization floors, see Nebius Form 6‑K and Q2 letter).
Spot/preemptible tiers can be 40-60% cheaper than on‑demand list rates (Nebius preemptible H100 $2.15 vs $3.85 on‑demand; CoreWeave spot H100 $2.46 vs derived $6.16 on‑demand). But spot capacity is interruptible and better suited to fault‑tolerant workloads.
Sales‑led GB300/B300/NVL72 purchases are often not on public rate cards. Vendors will route next‑gen Rubin/GB class systems through enterprise channels rather than self‑service pricing. If you need NVL72/GB300 at scale, ask for explicit availability windows and a delivery schedule, not a promise of “access.”
Per‑token models (Groq) reshape economics for inference. A low per‑token raster can beat GPU‑hour billing for high‑throughput inference, but you must benchmark with your model and input mix.

Workload fit, practical cheat sheet and negotiation levers

Frontier training (NVL72 / GB300 appetite), CoreWeave if you need validated NVL72 systems and an enterprise SLA. Nebius if you prioritize headline B300 access and will accept heavy contract mechanics. Negotiation levers: reserved capacity percentage, hardware uplift and refresh terms, and guaranteed delivery windows for NVL72 racks.
Mid‑size training clusters, Lambda for straightforward, self‑serve mid‑market clusters. Nebius if price sensitivity is high and you can structure prepayments. Negotiation levers: per‑GPU price bands by utilization, trial cluster credits, and a modest reserved term to lock price.
Budget experiments / spot training, Nebius or CoreWeave spot/preemptible tiers. Prioritize automated checkpointing and model resumption testing. Negotiate a short pilot with guaranteed minimum replenishment of spot capacity during the test window.
AMD evaluation, Crusoe (MI300X/MI355X on the rate card). Ask for a 7-14 day evaluation cluster with your training routine and a published spec sheet for the MI300X node used.
High‑volume production inference, Groq for per‑token economics on LPU endpoints. Nebius or Crusoe for GPU‑based inference depending on latency and regional coverage. Negotiate SLOs tied to percentile latency and a simple token‑to‑cost conversion example on your model during the RFP.

Vendor signals buyers and investors should watch

CoreWeave capex and backlog: guidance of $12.4-13.2B revenue for 2026 alongside a $35-39B capex guide is unusually large relative to revenue and deserves scrutiny. Ask vendors to show funded capex sources (debt, sale‑leaseback, customer prepayments). CoreWeave lists recourse debt and term loans on its balance page (CoreWeave Q2’26 release).
Nebius contract mechanics: the Nebius, Meta structure (March 2026) and other large deals show how prepaid dollars and utilization‑floor clauses can turn low list prices into durable cash flows (see Nebius Form 6‑K and investor letter).
Crusoe campus builds: Oracle/OpenAI and Microsoft campus projects change the vendor profile from spot supplier to long‑term capacity partner. Confirm milestones, financing sources, and construction loan commitments.
Groq licensing & partnerships: NVIDIA’s license of Groq inference IP and Groq’s entry into the NVIDIA Cloud Partner program materially reshape where Groq fits in customers’ stack. Validate cross‑support paths if you plan hybrid LPU plus NVIDIA GPU inference.

Risks and concrete vendor tests to run

Request historic realized pricing. Ask each vendor for anonymized, invoice‑level realized GPU‑hr pricing from three comparable customers over the past 12 months. This shows the gap between list and clearing price.
NVL72/GB300 availability test. Demand a 48-72 hour NVL72 benchmark window with your model and a written delivery window for additional units. Public bring‑up (CoreWeave) is encouraging, but production scale availability is the question.
Spot/preemptible resilience test. Run a 7‑day training job that tolerates preemption and measure checkpoint recovery time and cost per converged epoch across providers.
Per‑token proof for inference. For Groq, Nebius, or others offering per‑token billing, request a 7‑day production inference pilot with your traffic pattern so you can calculate tokens to dollars for your actual load.
Contract detail checklist to include in RFP: prepayment percentage and refund mechanics, utilization floor clauses, termination and hardware refresh terms, SLA percentile latency and penalties, and regional availability guarantees.

Key takeaways, quick Q&A

Which provider has the highest independent quality rating?

CoreWeave. SemiAnalysis’s ClusterMAX 2.0 lists CoreWeave as the only Platinum provider (ClusterMAX report, Nov 2025). Action: ask vendors for the specific ClusterMAX excerpts they rely on and test NVL72 performance with your model.
Who publishes the cheapest H100 on‑demand list price?

Nebius publishes H100 (HGX) at $3.85/GPU‑hr (prices checked Aug 21, 2026). Action: if you’re price‑sensitive, request Nebius’ realized pricing for a comparable contract size and a trial on preemptible capacity to validate end‑to‑end cost.
Who publishes B300 on‑demand pricing?

Nebius is the only provider in this set that publishes a B300 on‑demand price ($7.85/GPU‑hr as of Aug 21, 2026). Action: confirm geographic availability and whether that list price applies to self‑serve or requires a sales contract.
Which provider lists AMD MI300X/MI355X on their public rate card?

Crusoe. They publish MI300X on the rate card and MI355X via sales (Crusoe pricing page). Action: request an MI300X performance benchmark with your model and a short evaluation cluster contract.
How should buyers interpret published list prices?

As a starting point. Realized unit economics often depend on prepayments, reserved discounts, auctions, and bespoke contract terms, Nebius and CoreWeave examples show how list prices can understate or be reshaped by contract structure. Action: include clauses in the RFP that surface realized pricing and typical reserved‑discount ladders for the vendor.

Data & sources (select items checked Aug 21, 2026)

CoreWeave Q2 2026 earnings release and CoreWeave pricing/data‑center pages (CoreWeave Q2’26 release; CoreWeave pricing page).
Nebius Q2 2026 results and investor letter, Nebius pricing page, Nebius Form 6‑K (Meta agreement) (Nebius Q2’26 materials; Form 6‑K filings).
Lambda pricing page, Lambda press releases (Series E Nov 2025; $1B credit facility May 2026; Microsoft agreement Nov 3, 2025).
Crusoe Series E announcement (Oct 2025), Abilene campus announcements (Mar/Jun 2026), and Crusoe pricing page.
Groq newsroom items (Dec 2025 licensing news; Jun/Aug 2026 fundraising), Groq NVIDIA Cloud Partner announcement (Aug 12, 2026).
SemiAnalysis ClusterMAX 2.0 report (Nov 2025) for independent cluster ratings.
Media reporting and filings cited in vendor materials (CNBC, Bloomberg, TechCrunch, DCD; referenced where the vendor did not publish full details).

Buying GPU capacity in 2026 is not a commodity purchase. You’re buying capacity, roadmap access, financing structure, and contract design. Use the list prices to shortlist, then run the tests above and ask for invoice‑level realized pricing, NVL72 availability dates, sample SLOs, and a short pilot with your workloads. That’s where sticker rates meet real economics.